xAI launches Grok 4.7 at Grok 4.6’s price

• Grok 4.7 retains Grok 4.6’s $2-per-million input and $6-per-million output list prices, according to minchoi's score summary.
• Terminal-Bench 4.0 rose from 20.3% to 38.0%, a +17.7-point first-party result reported in ericzakariasson's model-card post.
• Independent testing puts the Artificial Analysis Intelligence Index at 44 to 46, while minchoi's score summary records xAI’s 40.4% to 46.3% CursorBench comparison.
• Cursor, Grok Build and the Grok API have day-one access, as SpaceXAI's availability post confirms.

The official launch post describes a larger base model and a longer reinforcement-learning run aimed at multi-hour work. Artificial Analysis’s evaluation found the biggest external gains in agentic knowledge work, alongside 81,000 output tokens per Intelligence Index task at xhigh effort.

• Cursor, Grok Build and the Grok API are live now, per SpaceXAI's availability post.
• API list prices remain $2 per million input tokens and $6 per million output tokens, as minchoi's score summary notes and the official launch post confirms.
• A faster variant doubles output speed and price, while xAI says the standard model keeps Grok 4.6’s speed, according to SpaceXAI's launch post.

First-party

• CursorBench 4.0: 40.4% → 46.3%, +5.9 points, in minchoi's score summary.
• Terminal-Bench 4.0: 20.3% → 38.0%, +17.7 points, per ericzakariasson's model-card post.
• HealthBench Professional: 48.5% → 56.7%, +8.2 points, per ericzakariasson's model-card post.
• Harvey Legal Agent Benchmark: 15.8% → 19.6%, +3.8 points, per ericzakariasson's model-card post.

Third-party evaluators

• Artificial Analysis Intelligence Index: 44 → 46, +2 points, in Artificial Analysis’s evaluation at xhigh effort.
• Artificial Analysis Coding Agent Index with Grok Build: 47 → 56, +9 points, in Artificial Analysis’s evaluation.
• AA-Briefcase, a long-horizon professional-work benchmark: 1,546 Elo → 1,657 Elo, +111 Elo, in Artificial Analysis’s evaluation.

Customer-reported

Cursor’s launch-day discussion asks users for long-session results rather than publishing a customer score.

Independent evaluation found uneven movement outside agentic work. Artificial Analysis’s evaluation reports a 3.7-point decline on AA-LCR and a 1.1-point decline on AutomationBench-AA.

• AA-Omniscience accuracy: 48% → 47%, -1 point, in Artificial Analysis’s evaluation, even as its hallucination rate fell from 34% to 29%.
• Output use per Intelligence Index task: 36,000 → 81,000 tokens, +125%, in Artificial Analysis’s evaluation. Those figures compare 4.6 at high effort with 4.7 at xhigh.
• CursorBench 4.0: Fable 5.1’s 51.8% → Grok 4.7’s 46.3%, -5.5 points, in xAI’s launch table.

• Base and training: xAI describes a new, larger base model plus a longer reinforcement-learning run on a harder mixture weighted toward tasks that take many hours. The launch post gives no parameter count.
• Agent harness: training to natively understand the Grok Bot harness is meant to improve conversational work and general knowledge tasks, according to SpaceXAI's model description.
• Context: the window remains 500,000 tokens, according to Artificial Analysis’s evaluation.
• Safety: xAI says 4.7 uses an entirely new safeguard stack, allowed 3.3% of risky dual-use cyber prompts through on its HackerBench v0.3 test, and topped its cited LatchBio biosafety benchmark at 62.4%.

minchoi’s examples roundup collects showcase posts rather than a standardized creative evaluation, but the early material is unusually centered on interactive output.

• SpaceXAI’s city-game comparison puts open-world city builds from 4.7 and 4.6 side by side.
• A Blender scene made with “almost no direction” appears in minchoi's Blender example, while minchoi's robot-arm animation attributes a robot-arm animation to HTML generated by 4.7.
• A complete playable game from a one-line prompt is the claim in minchoi's one-line game example.
• minchoi’s 70-hour claim describes an autonomous run on one goal lasting more than 70 hours.
• A Three.js cloth simulation appears in minchoi's cloth-simulation clip.